Sarvam AI witch-hunt: dev says 4,000 GPUs aren't enough to pretrain a model from scratch
cneuralnetwork · x · 2026-11-21
As Indian AI startup Sarvam AI faces a public 'witch-hunt' over its founders' rhetoric (e.g., calling India 'data colonies of the West'), a developer publicly defends the company and offers a blunt technical take: if you believe 4,000 GPUs are enough to pretrain a model from scratch, 'you don't know how this thing works.'
He argues the criticism is logically incoherent: a company spending its own money and resources to explore a new direction has done nothing wrong — comparing it to a YC-backed startup pivoting with revenue it earned, which nobody would call a scam. He invites Sarvam employees to DM him to speak anonymously, noting they 'troll but hardly answer any questions.'
Related event: Sarvam AI Drawn Into GPU Debate as Insider Claims B300 Training(3 posts)→
More from Companies & People
- Jensen Huang repeatedly interrupted Mistral CEO at panel: "You're not even speaking human" — mattturck · 2026-11-21
- Hugging Face CEO Clem Delangue: open source is key to fighting power concentration in AI — mervenoyann · 2026-11-21
- Shareholder letter urges NVIDIA board to cede Jensen Huang reinstatement decision to new committee — kevinnbass · 2026-11-21
- Kevin Bass defends Jensen Huang amid management-style controversy: 'Show me the results, not out-of-context stories' — kevinnbass · 2026-11-21
- Jack Rae, ex-Gemini pretraining lead, has left Meta — morqon · 2026-11-21
- AI trainer's pivot: models now do 'most things', so the bottleneck is the idea — beglen · 2026-11-21